CN110619642B

CN110619642B - Method for separating seal and background characters in bill image

Info

Publication number: CN110619642B
Application number: CN201910835331.XA
Authority: CN
Inventors: 王俊峰; 高琳; 唐鹏; 李征
Original assignee: Sichuan University
Current assignee: Sichuan University
Priority date: 2019-09-05
Filing date: 2019-09-05
Publication date: 2022-02-01
Anticipated expiration: 2039-09-05
Also published as: CN110619642A

Abstract

The invention discloses a method for separating a seal from background characters in a note image, which comprises the following steps of firstly, collecting the note image containing the seal to establish a note seal data set; then, training a target detection model based on a convolutional neural network by using a labeling data set; secondly, detecting and positioning a seal image area by the trained model; then, carrying out color space transformation on the extracted stamp image; separating the seal and the background characters in the image by the blind source separation of the digital image; and finally, carrying out image segmentation on the separated stamp and the background character image to obtain a final result image. The method has better robustness to complex conditions such as uneven illumination, noise interference and the like, has better universal applicability, is suitable for the seal and background characters with any color or shape, can accurately separate the seal and the background characters, simultaneously retains the information in the seal and the background character information, and improves the accuracy and reliability of bill character recognition.

Description

Method for separating seal and background characters in bill image

Technical Field

The invention belongs to the field of computer digital image processing, and particularly relates to a method for separating a seal from background characters in a bill image.

Background

The bills are transaction vouchers of enterprises or individuals in commercial activities, and the number of the bills is increased sharply along with the rapid development of economy in China. The financial data informatization management system which is generally applied at present provides great convenience for inquiry and management of bill information, and a considerable part of the bill information is acquired from paper bills. Traditional collection mode is through the manual completion of typeeing of financial staff, because the information quantity is huge, need drop into a large amount of manual works, simultaneously because the reliability of manual work typeeing can't be ensured, still need spend a lot of manpowers and carry out the later stage proofreading. With the further improvement of financial information management capability, higher requirements are also placed on the accuracy and the input efficiency of bill information input. By utilizing the digital image recognition technology, the bill characters can be quickly and accurately positioned and extracted, the bill information is obtained through character recognition, and the input is automatically completed, so that the work efficiency of information input is greatly improved, and the error risk caused by manual operation is reduced while the input of manpower and material resources is reduced.

The bills are generally stamped with special stamps of tax or financial departments, and the stamping positions of some stamps are not fixed, so that important information on the bills can be covered or overlapped, which causes serious interference to subsequent character recognition. Therefore, in the bill image recognition process, it is usually necessary to restore the information covered by the stamp and then perform recognition. The traditional method for removing the seal is to separate the seal from the bill characters by separating a color channel on the assumption that the seal and the bill characters have different colors. However, there may be many colors of the stamp, and the stamp with the same color may have a large deviation from the standard color due to the difference of the ink and the like, and it is often difficult to accurately define and quantify the color of the stamp. In addition, the stamp itself also contains text information, which is also needed by financial staff, and only removing the stamp cannot meet the actual requirement, so that the stamp and the background text in the image need to be recovered at the same time.

Disclosure of Invention

The invention aims to solve the technical problem of providing a method for separating a seal from background characters in a bill image, which can accurately separate the seal from the background characters, improve the accuracy and reliability of bill character recognition and provide effective data for subsequent bill character recognition.

In order to solve the technical problems, the invention adopts the technical scheme that:

a method for separating a seal from background characters in a bill image comprises the following steps:

step 1: denoising the acquired note image, marking the position and the size of the seal in the image, and establishing a note seal data set;

step 2: training a target detection model based on a convolutional neural network according to the labeled data set to obtain seal detection model parameters;

and step 3: detecting a note image to be separated by using a trained seal detection model, positioning the note image to a seal area in the note image, and extracting seal area data;

and 4, step 4: carrying out color space transformation on the extracted seal area to obtain a transformed image;

and 5: through the blind source separation of the digital image, the seal in the image after the transformation is separated from the background characters, which specifically comprises the following steps:

step 51: respectively removing the mean value of the seal areas of the three channels of hue, saturation and brightness, and subtracting the image mean value from the seal areas to enable the image pixel value mean value to be zero;

step 52: then, whitening processing is carried out on the image after the mean value is removed, and a whitened image is obtained;

step 53: separating the seal and the background characters from the whitened image by using an independent component extraction method;

step 6: and carrying out image segmentation on the separated seal and background characters, and removing the interference of background objects to obtain a final image.

Further, the step 1 specifically comprises:

step 11: carrying out Gaussian smoothing denoising processing on the bill image to obtain a denoised bill image sample;

step 12: marking the denoised bill image sample, marking the position coordinate of the seal, storing the marking information as a text, and establishing a data set together with the original bill image.

Further, the step 2 specifically comprises:

step 21: pre-training the convolutional neural network by using a public image data set to obtain initial parameters of a detection model;

step 22: and further training the detection model by using the established seal data set to obtain seal detection model parameters.

Further, the step 4 specifically includes:

step 41: converting the seal image from an RGB color space to an HSV color space;

step 42: and decomposing the color channel into three channels of hue, saturation and brightness.

Further, in step 6, the image is segmented using an OTSU adaptive threshold segmentation algorithm.

Compared with the prior art, the invention has the beneficial effects that: 1) the method has better robustness to common complex conditions such as uneven illumination, noise interference and the like; 2) the method has better universal applicability, and the seal and the background characters in any color or shape can be separated by applying the scheme of the invention; 3) by adopting a target detection model based on a convolutional neural network, all seal image areas in the bill image can be quickly and accurately positioned; 4) through blind source image separation, the seal and the background characters can be accurately separated, meanwhile, the information in the seal and the information of the background characters are kept, and the accuracy and reliability of bill character recognition are improved.

Drawings

FIG. 1 is a schematic flow diagram of the process of the present invention.

FIG. 2 is a schematic diagram of a stamp image extracted by the method of the present invention.

FIG. 3 is a schematic diagram of the method of the present invention for decomposing an image into three channel images of H (for hue), S (for saturation), and V (for brightness).

FIG. 4 is a schematic diagram of the segmentation result of the stamp image separated by the method of the present invention.

FIG. 5 is a diagram illustrating the segmentation result of the background text image separated by the method of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

As shown in fig. 1, the method for separating the stamp and the background text from the bill image comprises the following steps:

the method comprises the following steps: collecting a batch of bill images containing the seal, carrying out denoising pretreatment on the bill images, and filtering the bill images by using a Gaussian smoothing filter to remove Gaussian noise in the bill images; then, marking out external rectangles corresponding to all the seals in the note image, wherein the external rectangles comprise coordinates of the upper left corner of the rectangle, the width and the height of the rectangle; and storing the labeling result as a file in a text format, and forming a bill seal data set together with the original bill image. By removing the noise in the bill image, the interference caused by the noise in the subsequent processing is eliminated. The position of the seal region image is marked, so that a seal data set is conveniently established, and a seal detection model is trained.

Step two: and detecting the note image to be separated by using the trained seal detection model. Firstly, pre-training a network model by using a public data set ImageNet to obtain initial parameters of the model; and then converting the seal marking data into a format required by the training of the fast RCNN network model, and training the pre-trained network model. The initial parameters of the detection model are obtained through pre-training on the public data set, the model can have the capability of extracting the characteristics of a general target, the seal data set is used for training, the detection model can be rapidly transferred to the seal target on the seal data set with small sample amount, and the training efficiency is improved.

Step three: inputting the bill image into a trained detection model, selecting a rectangular region corresponding to a result with a confidence coefficient greater than 0.9 (the confidence coefficient value range is 0-1) as a seal image region according to a result output by the model, and extracting data of the regions as an independent seal image, as shown in fig. 2.

Step four: carrying out color space transformation on the extracted stamp image to obtain a corresponding transformed image; the concrete implementation steps are as follows: the original seal image is an RGB three-channel color image, and the seal image is converted into an HSV color space; the HSV color image is subjected to channel decomposition to obtain H, S, V grayscale images corresponding to the three channels, and each grayscale image is processed as an independent observation image (as shown in fig. 3) in the subsequent steps. After the color channels are decomposed, the image data of each channel can be analyzed independently. The image is decomposed into image channels with three different attributes of hue, saturation and brightness, and observation channel observation data are provided for subsequent blind source image separation. The adopted image blind source separation method requires that the number of observed images is not less than that of source images, and for stamp images, only one source is used, and the number of the observed images is increased through color space transformation and channel decomposition.

Step five: realizing blind source separation of the digital image through independent component analysis, and separating the seal in the seal image from background characters; the concrete implementation steps are as follows: respectively calculating the gray level average value of the image aiming at three images of the same seal, namely the gray level images corresponding to H, S, V channels, and then subtracting the image average value from each pixel in the image to obtain the average value of the pixel gray level value of the image, wherein the average value is zero; and carrying out whitening processing on the image subjected to the mean value removal to obtain a whitened image. The correlation among the image features is reduced through whitening processing, so that the image features have the same variance, and subsequent independent component extraction is facilitated; and separating the seal and the background characters from the whitened image by utilizing a FastICA independent component extraction algorithm.

The image is subjected to mean value removing processing and whitening processing, so that the subsequent independent component extraction process is simplified, and the convergence and stability of independent component extraction are improved.

Step six: and respectively carrying out image segmentation on the separated stamp image and the background character image, and segmenting the image by using an OTSU (over the Top) adaptive threshold segmentation algorithm to remove the interference of background objects, so as to obtain a final stamp image (shown in figure 4) and a final background character image (shown in figure 5).

Claims

1. A method for separating a seal from background characters in a bill image is characterized by comprising the following steps:

step 53: separating the seal and the background characters from the whitened image by utilizing a FastICA independent component extraction method;

step 6: and carrying out image segmentation on the separated stamp image and the background character image, and removing the interference of background objects to obtain a final image.

2. The method for separating the stamp from the background text in the bill image according to claim 1, wherein the step 1 is specifically as follows:

3. The method for separating the stamp from the background text in the bill image according to claim 1, wherein the step 2 is specifically as follows:

4. The method for separating the stamp from the background text in the bill image according to claim 1, wherein the step 4 is specifically as follows:

5. The method according to claim 1, wherein in step 6, the image is segmented using OTSU adaptive threshold segmentation algorithm.